Human Learning meets Machine Learning

1,200+ hours of piano practice

1 Introduction

1.1 What am I hoping to achieve with this project?

• predict how long it will take to learn a piece based on various features
• discover insights into my practice habits and find areas where I need to improve
• develop a recommender tool of piano pieces for others to select from based on factors of interest
• hopefully act as a source of inspiration/realistic piano progress for others who want to learn a musical instrument

Context: I started playing the piano in 2018 as a complete beginner and I’ve been tracking my practice time for around 2 and a half years. I now decided to put that to good use and see what interesting patterns I might be able to find.

1.2 Data collection

• imputed conservative estimations for the first 10 months of the first year (Jan ’18 to Oct ’18)
• everything from Dec ’18 onwards was tracked using Toggl, a time-tracking app/tool
• these times include my practice sessions (usually in short bouts of 30 minutes); piano lessons are excluded (usually 2-3 hours total per month)
• the Extract, Transform, Load script is available in the global.R file of this repo

Disclaimer: I am not affiliated with Toggl. I started using it a few years ago because it provided all the functionality I needed and loved its minimalistic design. The standard membership, which I use, is free of charge.

2 Key insights

2.1 Summary:

• identified various trends in average daily practice session, time of year, etc
• pieces could take me anywhere from around 4 hours to 40+ hours, subject to difficulty (as assessed by the ABRSM grade)
• the Random Forrest model was shown to be the most optimal model
  • Rsquared (0.585)
  • MAE - 5.9 hours
  • RMSE - 7.5 hours
  • Looking at the variability of errors, there is still a tendency to over-predict for pieces that took very little and under-predict for the more difficult ones. There could be two main reasons for this:
    • practicing an old piece in order to further improve (which naturally adds more practice time as I re-learn it)
    • learning easier pieces later on in my journey which means I will learn them faster than expected (based on my earlier data where a piece of a similar difficulty took longer)
• the most important variables were shown to be the number of length of the piece, standard of playing (performance vs casual) and experience(cumulative hours before first practice session on each piece)

knitr::opts_chunk$set(
    echo = FALSE, # show all of the code
    tidy = FALSE, # cleaner code printing
    size = "small", # smaller code
    
    fig.path = "figs/",# where the figures will end up
    out.width = "100%",

    message = FALSE,
    warning = FALSE
    )

3 Exploratory Data Analysis (EDA)

3.1 Piano practice timeline

3.2 How long did I practice per piece?

Based on the level at the time and the difficulty of the piece, we can see that each piece took around 10-30 hours of practice.

3.3 How consistent was my practice?

Generally, I’ve done pretty well to maintain a high level of consistency with the exception of August/December. This is usually where I tend to take annual leave.

3.4 Was there a trend in my amount of daily average practice?

We can see that my practice time was correlated with the consistency, where the average session was much shorter in the months I was away from the piano. There’s also a trend where my practice close to an exam session was significantly higher than any other time of the year. Can you spot in which month I had my exam in 2019? What about the end of 2020?

average practice length per month includes the days in which I did not practice

overall

Year on Year

Similar trends as before are apparent where my average daily session is longer before the exams than any other time in the year and a dip in the months where I usually take most of my annual leave. I really do need to start picking up the pace and get back to where I used to be.

3.5 Did COVID significantly impact my practice time?

graph

Despite a similar median, we can see that the practice sessions were less likely to be over 80 min after COVID. We can test if this was a significant impact with a t-test.

skewness assumption

Given the extremely low p-value, the Shapiro-Wilk normality test implies that the distribution of the data is significantly different from a normal distribution and that we cannot assume the normality. However, we’re working with the entire population dataset for each class and thus, unlike the independence of data, this assumption is not crucial.

Shapiro-Wilk normality test

group	statistic	p.value	method
After COVID	0.9607325	3e-07	Shapiro-Wilk normality test
Before COVID	0.9549818	0e+00	Shapiro-Wilk normality test

equal variances assumption

We can see that with a large p value, we should fail to reject the Null hypothesis (Ho) and conclude that we do not have evidence to believe that population variances are not equal and use the equal variances assumption for our t test

Levene’s test

statistic	p.value	df	df.residual
0.0410026	0.8395891	1	732

t-test

My practice sessions post-COVID are significantly shorter than those before the pandemic. This might be surprising, given that we were in the lockdown most of the time. However, I’ve been spending my time doing a few other things such as improving my technical skillset with R (this analysis wouldn’t have been possible otherwise) and learning italian.

.y.	group1	group2	n1	n2	statistic	df	p	p.signif
Duration	Before COVID	After COVID	433	301	3.319481	732	0.000947	***

3.6 What type of music do I tend to play?

by genre

by composer

by piece

3.7 Relation between difficulty and number of practice hours

ABRSM grade

Simplified, ABBRSM grades are a group of 8 exams based on their difficulty (1 - beginner to 8 - advanced). There are also diploma grades but those are extremely advanced, equivalent to university level studies and out of the scope of this analysis.

More information can be found on their official website at https://gb.abrsm.org/en/exam-support/your-guide-to-abrsm-exams/

level

A further aggregation of ABRSM grades; this is helpful given the very limited dataset within each grade and much easier on the eye. This is an oversimplification but they’re classified as: * 1-5: Beginner (1) * 5-6: Intermediate (2) * 7-8: Advanced (3)

3.8 What about the piece length?

3.9 Learning effect - do pieces of the same difficulty become easier to learn with time?

We can spot a trend where the time required to learn a piece of a similar difficulty (ABRSM Grade) decreases as my ability to play the piano increases (as judged by cumulative hours of practice). We should keep this in mind and include it as a variable into our prediction model.

3.10 Does “pausing” a piece impact the total time required to learn it?

How do we differentiate between pieces that we learn once and those that we come back to repeatedly? Examples could include wanting to improve the playing further, loving it so much we wanted to relearn it, preparing it for a new performance, etc.

As anyone that ever played the piano knows, re-learning a piece, particularly after you “drop” it for a few months/years, results in a much better performance/understanding of the piece. I definitely found that to be true in my experience, particularly with my exam pieces.The downside is that these pieces take longer to learn.

3.11 Repertoire

Repertoire

Project	Duration	Genre	ABRSM	Level	Standard	Length	Experience	Break	Started
Elton John - Rocket man	47	Modern	7	Advanced	Performance	4.0	1130	No	2020-12-08
Schumann - Träumerei	14	Romantic	7	Advanced	Average	3.0	1087	No	2020-11-09
Mozart - Allegro (3rd movement) K282	28	Classical	6	Intermediate	Average	3.3	1081	Yes	2020-11-05
Ibert - Sérénade sur l’eau	10	Modern	6	Intermediate	Performance	1.7	1038	No	2020-09-24
Kuhlau - Rondo Vivace	24	Classical	6	Intermediate	Average	2.2	1014	No	2020-08-03
C. Hartmann - The little ballerina	21	Romantic	6	Intermediate	Performance	2.0	998	No	2020-07-14
Schumann - Lalling Melody	5	Romantic	1	Beginner	Average	1.3	981	No	2020-06-28
Schumann - Melody	4	Romantic	1	Beginner	Average	1.0	972	No	2020-06-20
Clementi - Sonatina no 3 - Mov 2	3	Classical	1	Beginner	Performance	1.0	952	No	2020-06-04
Clementi - Sonatina no 3 - Mov 3	20	Classical	4	Beginner	Performance	2.0	952	No	2020-06-04
Chopin - Waltz in Fm	27	Romantic	6	Intermediate	Performance	2.0	895	Yes	2020-04-18
Clementi - Sonatina no 3 - Mov 1	30	Classical	4	Beginner	Performance	2.7	877	No	2020-04-07
Schumann - Kinderszenen 1	10	Romantic	5	Intermediate	Average	2.0	855	No	2020-03-25
Bach - Prelude in G from Cello Suite No 1	25	Baroque	5	Intermediate	Average	2.5	788	No	2020-02-04
Georg Böhm - Minuet in G	7	Baroque	1	Beginner	Average	1.0	780	Yes	2020-01-27
Bach - Invention 4 in Dm	21	Baroque	5	Intermediate	Performance	1.7	777	No	2020-01-25
Chopin - Contredanse in Gb	23	Romantic	6	Intermediate	Performance	2.2	762	No	2020-01-16
Bach - Minuet in Gm - 115	7	Baroque	1	Beginner	Average	1.3	750	No	2020-01-07
Bach - Minuet in G - 114	4	Baroque	1	Beginner	Average	2.0	726	No	2019-12-06
Elton John - Your song (Arr Cornick)	36	Modern	5	Intermediate	Performance	3.3	713	No	2019-11-21
Poulenc - Valse Tyrolienne	17	Modern	5	Intermediate	Performance	1.7	562	No	2019-09-02
Bach - Prelude in Cm - 934	25	Baroque	5	Intermediate	Performance	2.4	536	No	2019-08-15
Schumann - Volksliedchen	10	Romantic	2	Beginner	Average	1.8	501	No	2019-07-01
Haydn - Andante in A	39	Classical	5	Intermediate	Average	2.8	468	Yes	2019-06-08
Schumann - Remembrance	34	Romantic	5	Intermediate	Performance	2.2	422	Yes	2019-04-28
Bach - Minuet in G - 116	8	Baroque	1	Beginner	Average	2.0	361	Yes	2019-03-04
Bach - Invention 1 in C	27	Baroque	5	Intermediate	Performance	1.7	350	Yes	2019-02-22
Chopin - Waltz in Am	26	Romantic	4	Beginner	Performance	2.5	305	Yes	2019-01-07

4 Modelling

Question: How long would it take to learn a piece based on various factors?

4.1 detect outliers

Given the very limited data at the advanced level (Grade 7 ABRSM), those two pieces will be removed. One is an extreme outlier as well which will significantly impact our models.

4.2 missing values

There are no missing values in the modelling dataset following the ETL process.

4.3 Feature engineering

• categorical:
  • ABRSM grade: 1 to 8
  • Genre: Baroque, Classical, Romantic, Modern
  • Break: learning it continuously or setting it aside for a while (1 month minimum)
  • Standard of practice: public performance or average (relative to someone’s level of playing)
• numerical:
  • Experience: total hours practiced before the first practice session on each piece
  • piece Length: minutes

4.4 Pre-processing

Let’s use some basic standardisation offered by the caret package such as centering (subtract mean from values) and scaling (divide values by standard deviation).

4.5 Resampling

Given the small size of the dataset, bootstrapping resampling method will be applied.

4.6 Model selection

# set number of clusters 
clusters <- 4

# run them all in parallel
cl <- makeCluster(clusters, type = "SOCK")
 
# register cluster train in paralel
registerDoSNOW(cl)

# train models
model <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                  data = model_data,
                  method = "ranger",
                  tuneLength = 100,
                  trControl = train.control)


model2 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "lmStepAIC",
                tuneLength = 100,
                trControl = train.control)


model3 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "lm",
                tuneLength = 100,
                trControl = train.control)

model4 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "ridge",
                tuneLength = 100,
                trControl = train.control)

model5 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "rf",
                tuneLength = 100,
                trControl = train.control)

model6 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "gbm",
                tuneLength = 100,
                trControl = train.control)

model7 <- train(Duration ~ ABRSM + Genre + Length + Cumulative_Duration + Break + Standard,
                data = model_data,
                method = "pls",
                tuneLength = 100,
                trControl = train.control)
 
# shut the instances of R down
stopCluster(cl)

# compare models
model_list <- list(ranger = model, lmStepAIC = model2, lm = model3, ridge = model4, rf = model5, gbm = model6, pls = model7)

model_comparison <- resamples(model_list)

# learning curves to indicate overfitting and underfitting
# hyper parameters 
# https://topepo.github.io/caret/model-training-and-tuning.html#model-training-and-parameter-tuning
# https://topepo.github.io/caret/random-hyperparameter-search.html

We chose the Random Forest model as it was the best performing model. It is known as a model which is:

• not very sensitive to outliers
• good for non-linearity
• variable importance can be biased if categorical variables have few levels (toward high levels) or are correlated

## 
## Call:
## summary.resamples(object = model_comparison)
## 
## Models: ranger, lmStepAIC, lm, ridge, rf, gbm, pls 
## Number of resamples: 25 
## 
## MAE 
##               Min.  1st Qu.   Median     Mean  3rd Qu.      Max. NA's
## ranger    3.216848 4.546412 5.226721 5.350732 6.185954  8.074644    0
## lmStepAIC 4.059373 6.035128 6.625829 7.373961 8.656703 12.293508    0
## lm        3.457466 6.104998 6.831580 7.407868 8.107952 16.638647    0
## ridge     3.514506 5.195608 5.904643 5.936369 7.002558  7.751690   10
## rf        2.867137 5.123509 5.592698 5.959335 6.989707  8.722035    0
## gbm       4.593238 6.146348 6.951988 7.261349 8.076775 16.990824    0
## pls       4.460543 5.041959 5.930350 5.848360 6.295627  7.970644    0
## 
## RMSE 
##               Min.  1st Qu.   Median     Mean   3rd Qu.      Max. NA's
## ranger    4.531594 5.821475 6.907062 6.876131  7.614550 10.521156    0
## lmStepAIC 4.678084 7.290730 8.706652 9.067898 10.179064 15.976038    0
## lm        3.979872 8.065115 8.705383 9.404728 10.412562 19.451739    0
## ridge     4.989087 6.206758 6.862180 7.045983  7.854493  8.687658   10
## rf        3.394308 6.640020 7.115649 7.546307  8.461888 11.065527    0
## gbm       5.906910 7.466537 8.041088 8.843606  8.994092 21.377506    0
## pls       5.322380 6.303446 7.096138 7.043631  7.345444  9.624458    0
## 
## Rsquared 
##                   Min.   1st Qu.    Median      Mean   3rd Qu.      Max. NA's
## ranger    0.2598148174 0.5915672 0.6861549 0.6395133 0.7899287 0.8801139    0
## lmStepAIC 0.0005882529 0.4001020 0.5735153 0.5051570 0.6597403 0.8997367    0
## lm        0.0019980780 0.2879700 0.5871362 0.4965277 0.6363397 0.8958706    0
## ridge     0.3743315404 0.6559275 0.6807293 0.6828744 0.7764227 0.8621330   10
## rf        0.1341300382 0.5248279 0.6068158 0.5858350 0.7012294 0.8842650    0
## gbm       0.0669333944 0.3848695 0.5685371 0.5338006 0.6661362 0.8728232    0
## pls       0.3579501758 0.6342263 0.6727956 0.6723318 0.7638235 0.8898141    0

Based on our regression model, it does not look like we have significant multicollinearity between the full model variables so we can continue as it is.

Variance Inflation Factor (VIF)

names	VIF
ABRSM6	4.0
Cumulative_Duration	3.9
ABRSM5	3.5
ABRSM4	3.4
GenreClassical	2.5
Length	2.4
StandardPerformance	2.4
BreakNo	2.1
GenreRomantic	1.9
ABRSM2	1.7
GenreModern	1.7

4.6.1 Actuals vs Predictions

4.6.2 Residual distribution

We can see that the residuals are mostly situated around 0.

4.6.3 Actuals versus Residuals

Looking at the variability of errors, there is still a tendency to over-predict for pieces that took very little and under-predict for the more difficult ones. There could be two main reasons for this:

• practicing an old piece in order to further improve (which naturally adds more practice time as I re-learn it)
• learning easier pieces later on in my journey which means I will learn them faster than expected (based on my earlier data where a piece of a similar difficulty took longer)

4.7 Linear Regression (LR) or Random Forest (RF)?

We can see that the Random Forest performed significantly better than the Linear Regression model. This isn’t surprising since there might be non-linear trends within the data, and RFs are known to be more accurate.

Model 1 vs model 2

estimate	statistic	p.value	parameter	conf.low	conf.high	method	alternative
1.858421	3.01994	0.0059183	24	0.5883319	3.12851	One Sample t-test	two.sided

4.8 How many predictors did the most optimal model have?

4.9 What were the most important variables?

We can now see that the most important variables seemed to be the length of the piece, my experience prior to starting a piece and time difficulty of the piece. These were also confirmed by the linear regression model.

5 Limitations

• very limited data which did not allow for a train/test split; however, a bootstrap resampling method is known to be a good substitute
• biased to one person’s learning ability (others might learn faster or slower)
• on top of total hours of practice, quality of practice is a significant factor which is not captured in this dataset
• very difficult to assess when a piece is “finished” as you can always further improve on your interpretation
• not all pieces had official ABRSM ratings and a few had to be estimated; even for those that do have an official rating, the difficulty of a piece is highly subjective to each pianist and hard to quantify with one number
• memorisation might be a confounding variable that was not accounted for; sometimes there’s an effort to practice a bit for longer just to memorise without an improvement in performance

6 Hardest things about this analysis:

• the Extract-Transform-Load process - clean the “dirty data” and find creative ways to input the data on the front end of the app to make it reporting friendly on the back-end (with all the variables such as Genre, Type of practice, Composer and Piece name, tag pieces as “work in progress” etc)
• automate ways to differentiate between pieces that I came back to vs pieces I only studied once
• work with very limited data

7 Interactive application:

• you can find an interactive display of this presentation, as well as the model in production at the following link
• https://peterhontaru.shinyapps.io/piano-practice-prediction/

8 What’s next?

• keep practicing, gather more data and refresh this analysis + update the model
• add a recommender tab to the shiny dashboard where people could be recommended a piece based on specific features
• connect to the Toggl API for live updates